Research On Point Cloud Registration Algorithm With Scene Classification

With the rise of RGB-D data acquisition system,indoor scene reconstruction using RGB-D data has become a hot spot in the field of 3D reconstruction.As a key part,point cloud registration technology has received extensive attention and research.In this paper,aiming at the point cloud registration problem between key frames in sparse mapping,by introducing the scene classification method,the coarse registration algorithm is studied combining geometric information and photometric information,and based on the coarse registration,the fine registration algorithm is further studied,which achieves accurate and robust transformation estimation between key frames.The main research contents of this paper are as follows:1.The RGB-D data acquisition system is introduced,the work principle of the system is analyzed and the system calibration experiment is designed and completed.The scene point cloud acquisition experiment is conducted.The preprocessing of the scene point cloud is realized,including data organization,adaptive downsampling,outlier elimination and normal estimation.2.The point cloud coarse registration algorithm based on improved random sample consensus(RANSAC)is studied.The key points are detected,described and matched separately from point cloud data and color image,and the adaptive combination of geometric correspondences and photometric correspondences is realized according to the scene category.An improved RANSAC is proposed,and robust and effective initial transformation estimation is achieved.3.The scene classification algorithm is studied.Scene feature is encoded with local binary pattern,a binary validity template is proposed,and effective extraction of the scene feature vector is achieved.The classifier is constructed using three-class support vector machine,and the training and testing of the classifier is completed.4.The point cloud fine registration algorithm based on iterative closest point(ICP)is studied.The popular ICP algorithm framework is analyzed.According to the algorithm framework,an improved ICP algorithm is proposed.Through the adjustment and the improvement of matching correspondences and optimization solution,accurate fast and stable fine registration is achieved.